Consumer Behaviour and Revealed Preference - …uctp39a/blundell-short-course-2017-lecture3.pdf ·...

Consumer Behaviour and Revealed PreferenceHow Revealing is Revealed Preference?

Richard Blundell (UCL and IFS) Northwestern University

Short Course November 2017

Richard Blundell () Consumer Behaviour & Revealed Preference Short Course November 2017 1 / 89

This lecture will focus on three aspects of RP research:

I. Testing Rationality using Revealed Preference

I Afriat-VarianI Experiments, Observational Data and the SMP idea

II. Using RP to Bound Counterfactual Demand Responses

I Using Nonparametric Expansion PathsI Unobserved Heterogeneity and Quantile Demands

III. Rationality and Taste Change

I Identifying Taste/ Quality Change: tobacco, environmental badsI Intertemporal Preferences and Information

Background references in the intro lecture and on my website.

I Afriat-VarianI Experiments, Observational Data and the SMP ideaII. Using RP to Bound Counterfactual Demand Responses

Rationality and Revealed Preference: Introduction

There are (at least) two key criticisms of the empirical application ofrevealed preference theory to consumer behaviour:I when it does not reject, it doesn’t provide precise

counterfactual predictions; andI when it does reject, it doesn’t help us characterize the nature

of irrationality or the degree/direction of changing tastes.

In this lecture we will argue that recent developments in themicroeconometric application of revealed preference have renderedthese criticisms unfounded.

Modern RP analysis takes a nonparametric approach.

To quote Dan McFadden: “parametric models interpose an untidyveil between econometric analysis and the propositions of economictheory”

The aim of this lecture is to “lift ‘McFadden’s’untidy veil”!

How are preferences revealed?

I Inequality restrictions from revealed preference are used to assessrationality and to improve the estimation of counterfactual demandresponses.

I Particular attention is given to application to observational data:nonseparable unobserved heterogeneity and endogeneity.

I New insights are provided about the price responsiveness and thedegree of rationality, especially across different income and educationgroups.

The analysis extends to....

General choice models...I Collective choiceI HabitsI Intertemporal choiceI Characteristics models

And ‘Beyond’...I Hyperbolic discountingI Choice under uncertaintyI Consideration setsI Reference-dependent choice...

The analysis extends to....

General choice models...I Collective choiceI HabitsI Intertemporal choiceI Characteristics models

And ‘Beyond’...I Hyperbolic discountingI Choice under uncertaintyI Consideration setsI Reference-dependent choice...

1. Nonparametric Revealed Preference

Observe (a sample analog of) demands (continuous case), or choiceprobabilities (discrete case), and ask: Can the observable choices, or choiceprobabilities, be rationalized as an outcome of optimisation?

The aim is (i) to devise a powerful test of RP conditions and (ii) to estimatedemand counterfactuals for some new budget using only the theoretical(shape) restrictions implied by the optimising framework.

Nonparametric RP relies on (preference) orderings avoiding parametricrestrictions on the form of utility. In general, we can only derive setidentification for counterfactual demands.

- Focus mainly here on continuous choice models, following work with Browningand Crawford (BBC (2003,2008), and with Kristensen and Matzkin (BKM (2014,2017), and also with Horowitz and Parey (BHP (2013, 2016).

- Should consider discrete choice models, following recent work of Kitamura andStoye (2017) focussing on the Axiom of Stochastic Revealed Preference (ASRP)from McFadden and Richter (1991), McFadden (2005), and Manski (2012).

1. The consumer problem

Assume every consumer is characterised by observed and unobservedheterogeneity (h, ε) and responds to a given budget B(p, x), with aunique, positive J−vector of demands

q = d(p, x ,h, ε).

Demand functions: RK++ → RJ++,satisfy adding-up: p′q = x for all pricesand total outlays x ∈ R ; ε ∈ RJ−1, J − 1 vector of unobservableheterogeneity. Assume ε⊥x | h, for now.The environment is described by a continuous distribution of q, x and ε, fordiscrete types h,

will often suppress observable heterogeneity h.

For discrete prices (finite set of markets), the demand curve for given pricesdefines the expansion path (Engel curve) for consumer (h, ε) as their totalbudget x (income) is varied:

q = g(x ;h, ε),

this plays a central role in RP analysis of consumer demand.Richard Blundell () Consumer Behaviour & Revealed Preference Short Course November 2017 7 / 89

1. The consumer problem (cont...)

One key assumption in first generation studies was (additive) separability ofε. In the non-separable case we will assume conditions on preferences thatensure invertibility in ε, equivalent to monotonicity for the scalarheterogeneity case when J = 2.

I Application: illustrate importance of flexibility in price responses acrossthe income distribution using gasoline demand, BHP (2013, 2016).

Let’s first abstract from heterogeneity and examine Afriat’s Theorem for asingle consumer, and the construction of support sets for counterfactualdemands under new budgets.

Afriat’s Theorem

Assume consumer faces a finite number of budgets (markets) and index thebudget sets as Bt = B(p(t), x), t ∈ T .

The Afriat-Diewert-Varian Theorem allows us to characterise ‘well behaved’preferences through a set of inequalities on observed price and quantityvectors (pt ,qt ) across markets t ∈ {1, ...,T} .

Provides the basis for a test of rationality which generalises to manyalternative rationality concepts, for both observational and experimentaldata.

Afriat’s Theorem (cont...)

Afriat’s Theorem: The following statements are equivalent:

A. there exists a utility function u (q) which is continuous,non-satiated and concave which rationalises the data {pt ,qt}t=1,...,TB1. there exist numbers {Ut ,λt > 0}t=1,...,T such that

Us ≤ Ut + λtp′t (qs − qt ) ∀ s, t ∈ {1, ...,T}

B2. the data {pt ,qt}t=1,...,T satisfy the Generalised Axiom ofRevealed Preference (GARP).

Definition: A dataset {pt ,qt}t=1,..,T satisfies GARP if and only if wecan construct relations R0,R such that(i) for all t, s if ptqt ≥ ptqs then qt R0 qs ;(ii) for all t, s, u, . . . , r , v , if qt R0 qs , qs R0 qu , . . . , qr R0 qv thenqt R qv ;(iii) for all t, s , if qt R qs , then psqs ≤ psqt .

Condition (i) states that the quantities qt are directly revealed preferred over qsif qt was chosen when qs was equally attainable.Condition (ii) imposes transitivity on the revealed preference relation R .

Condition (iii) states that if a consumption bundle qt is revealed preferred to aconsumption bundle qs , then qs cannot be more expensive then qt .

Figure 1a: illustrates a simple RP rejection.

Figure 1a

Figure 1a: Rejection

Support Sets and Bounds on Demand Responses

Suppose we observe a set of demand vectors {q1,q2, ...qT } which record thechoices made by a consumer when faced by the set of prices {p1,p2, ...pT } .� new price vector p0 with total outlay x0, budget B0(p0, x0).� ‘best’support set SV (p0, x0) for q (p0, x0) is given by:{

q0 :p′0q0 = x0, q0 ≥ 0 and

{pt ,qt}t=0...T satisfies GARP

S (p0, x0) is the identified set of demand responses for p0, x0, with properties:(1) S (p0, x0) is non-empty iff the data set {pt ,qt}t=1,...T satisfies GARP.(2) If the data set {pt ,qt}t=1...T satisfies GARP and p0 = pt for some t thenS (p0, x0) is the singleton {qt}.(3) S (p0, x0) is convex.

Illustrated for the two dimensional case in Figure 1b:

Figure 1b: The ‘Varian’Support Set with GARP

( )00 , xSV p1q

( )11, xp

( )22 , xp

( )00 , xp

Extensions: (dynamic extensions later in slides)

1 Homothetic and weak separability (and conditional demands)2 Characteristics models3 Non-unitary models and altruism

All the ‘first generation’applications either:

follow individuals in short/small panels (and in experimental settings),ortreat unobserved heterogeneity as additive and work with conditionalmean models.

Illustrate the latter using a Kernel regression application to theimproved bounds.

Also generalise to nonseparable unobserved heterogeneity.

Weak Separability

Partition our data into two sets of goods and prices{{p1t ,q

},{p2t ,q

}}t=1,...,T

A utility function is separable in the group 1 goods, if

{q1,q2} �{q1∗,q

2}⇐⇒ {q1,q2#} � {q1∗,q2#}for all q1,q1∗,q2 and q2#.That is preferences within group 1 are independent of the composition ofgroup 2.The functional representation is that a utility function u is (weakly)separable in the group 1 goods if we can find a "subutility function" v

and a "macro function" w(v ,q2) strictly increasing in v such that:

u(q1,q2

)= w(v(q1),q2).

Dimension reduction and two-stage budgeting.

Weak Separability

Theorem (weak separability). The following conditions are equivalent:(1) there exists a weakly separable, concave, monotonic, continuousnon-satiated utility function that rationalises the data;(2) there exist numbers {Vt ,Wt ,λt > 0, µt > 0}t=1,...,T thatsatisfy:

Vs ≤ Vt + µtp1′t

(q1s − q1t

)Ws ≤ Wt +

λtµt(Vs − Vt ) + λtp2′t

(q2s − q2t

)(3) the data

{p1t ,q1t

}t=1,...,T and

{1/µt ,p

2t ,Vt ,q2t

}t=1,...,T satisfy

GARP for some choice of {1/µt ,Vt}t=1,...,T that satisfies the Afriatinequalities.

An easier strategy is to check for GARP for all goods and also for the weaklyseparable sub-set.

Characteristics and Revealed Preference

I Consumer choice model is extended to

maxqV (z) subject to z = F(q) and p′q ≤ x ,q ≥ 0.

I Blow, Browning and Crawford (REStud, 2006), extend the set of RPinequalities in BBC to the linear characteristics model, where z = A′q.

I Underlying characteristics are not necessarily observed so a harderidentification problem.I They use scanner panel data. But otherwise identical approach.I Note that GARP has to be satisfied on the original set of goods.

Collective Models

� Collective models - families versus individuals.

Data is typically for families, and may record some assignable and exclusivegoods.

Individual labor supplies are often taken as examples.

Cherchye, De Rock and Vermeulen (2011, ...) extend the analysis presentedin this lecture to the collective choice case. Very recently to non-corporativemodels as well.

A pair of utility functions UA(qA,Q) and UB (qB ,Q) for two members Aand B who consume private goods (qA,qB ) and public goods Q.Observed demands satisfy collective rationalisation (CARP) if inequalitieshold over “personalised”quantities.

Not all personalised quantities are observed so a harder identificationproblem. But otherwise identical approach.

Nice applications to family labour supply.

Altruism

� Models of Altruism - ‘rational’altruistic preferences.

Andreoni and Miller (Ecta). Adapt measures in an experimental design toinclude payments to self and payments to others, U(πs ,πo ;γ), where γare the observable attributes of the game.

Extended set of RP inequalities.

Implementing Afriat’s Theorem

Given some data {pt ,qt}t=1,...,T to check for consistency with the theorywe can either

1 Determine whether there exist numbers {Ut ,λt > 0}t=1,...,T suchthat

Us ≤ Ut + λtp′t (qs − qt ) ∀ s, t ∈ {1, ...,T}or,

2 Determine whether the data satisfy GARP.

Suppose we have just two market observations {p1,p2;q1,q2}.Then the Afriat Inequalities

Us ≤ Ut + λtp′t (qs − qt ) and λt > 0, ∀ s, t ∈ {1, 2}

can be written as

0 0 0 01 −1 0 −p′2 (q1 − q2)−1 1 −p′1 (q2 − q1) 00 0 0 00 0 −1 00 0 0 −1

U1U2λ1λ2

0000−ε−ε

where ε is an arbitrarily small constant, or

Ax ≤ b

Applying Afriat’s Theorem

In essence we are asking whether there exist a solution to a set of linearinequalities. This is a linear programming problem and Dantzig’s "simplexalgorithm", can determine whether or not there is a feasible solution in afinite number of steps.

In general checking for consistency requires a linear program with 2Tvariables and T 2 constraints.

The fact that the number of constraints rises as the square of the number ofobservations can makes this condition computationally demanding inpractice for very large datasets.

Condition B2 (GARP) is sometimes more effi cient. This requires us tocompute the transitive closure of a finite relation. That is certainly a finiteproblem and Warshall (1962) gives a solution in T 3 steps. It is very easy toimplement, see BBC (2003), for example.

1. Can we improve on the Afriat ‘test’of rationality?

Clearly not without further assumptions, information or a change in theexperimental design.

We have seen that the Afriat-Diewert-Varian Theorem allows us tocharacterise ‘well behaved’preferences through a set of inequalities onobserved behaviour (pt ,qt )

Provides a test of rationality

Generalises to many alternative rationality concepts

Data: Both Observational and Experimental

Start here by asking if there is a best experimental design for testing RP?

Recall the simple RP rejection in: Figure 1a:

Figure 1a: a ‘rejection’region

Figure 1a: An uninformative budget

B(p3, x3)

2. Intersection Demands and Improving Bounds

Define sequential maximum power (SMP) path

{xs , xt , xu , ...xv , xw } = {p′sqt (xt ),p′tqu(xu),p′vqw (xw ), xw }

Proposition 2.1 (BBC, 2003): Suppose that the sequence

{qs (xs ) ,qt (xt ) ,qu (xu) ...,qv (xv ) ,qw (xw )}

rejects RP. Then SMP path also rejects RP. �This result has been used in the design of RP experiments and also extendedthis result to models of collective choice, habits, in the referenced papers...

Key idea for observational data: - use expansion paths (Engel curves) tomimic the experimental design.

Observe consumers across a finite set of markets (in each market they facethe same relative prices). Using expansion paths qt (x) (Engel curves) foreach market t, we are able to generate the SMP path. See Fig 2a.

Figure 2a: Using Expansion Paths

Bounds on Demand Responses Using Engel Curves

The expansion paths (Engel Curves) {qt (x)}t=1,..T , define intersectiondemands qt (xt ) by p′0qt (xt ) = x0.

The set of points that are consistent with observed expansion paths andutility maximisation is given by the support set:

S (p0, x0) ={q0 :

q0 ≥ 0, p′0q0 = x0{p0,pt ;q0,qt (xt )}t=1,...,T satisfy GARP

}The support set S (p0, x0) that uses expansion paths and intersectiondemands defines e-bounds on demand responses

S (p0, x0) is the identified set for the parameter q(p0, x0).

Proposition 2.2 (BBC2): the set is sharp and is convex, refined in BBCDV(AEJ-Micro, 2015).

See Figures 2 b,c,d

Figure 2b: The ‘Varian’Support Set with RP

( )00 , xSV p1q

( )11, xp

( )22 , xp

( )00 , xp

Figure 2c. Support set with Expansion Paths

( )x1q( )11,xp

( )22,xp

( )x2q

( )22~xq

( )11~xq

( )00 , xS p

( )00,xp

Figure 2d: Support Set with Many Markets

( )x1q( )11, xp

( )33 , xp

( )x2q

( )x3q

( )33~xq

( )22~xq

( )11~xq

( )00 , xS p

aqmaxaqmin

( )00 , xp

( )22 , xp

Transitivity

Transitivity, like symmetry, adds nothing in the two good case.Many good application to UK FES diary records on foods, servicesand some categories of other goods, BBC (Ecta, 2008).Implement the SMP idea using average local Engel curves for eachmarket estimated by nonparametric regression.

I Engel curves can be quite nonlinear (see QUAIDS and otherreferences).I Test rationality through RP inequality restrictions based onintersection demands.I Findings: periods of time for certain demographic groups for whichthe RP restrictions cannot be rejected.I Use restricted Engel curves to estimate bounds on counterfactualdemand responses.

Assume additive unobserved heterogeneity on Engel curves, BCK(2007) account for endogeneity.Also construct (and estimate) bounds on welfare costs of prices (tax)changes....Richard Blundell () Consumer Behaviour & Revealed Preference Short Course November 2017 34 / 89

Bounds on indifference surface and cost of living

- provide bounds on compensating variations.

Note that as the data becomes dense

- the RP test for consistency becomes more demanding

- the bounds on indifference curves become tighter

- the bounds on demand responses become tighter.

If the data become perfectly dense (effectively an infinite dataset) wehave the indifference curve map and demand curves themselves.

In this case the RP conditions become equivalent to the usualintegrability conditions - the Slutsky condition and homogeneity.

Will give some examples of the nonparametric implementation of Slutskycondition for this case.

Rationality and Revealed Preference:

Summary so far....

Inequality restrictions from revealed preference used

I to test rationality through inequality restrictions, andI to provide nonparametric estimates of bounds on counterfactualdemand responses.

The remainder of this lecture will- focus on unobserved heterogeneity with some examples- formalise the notion of taste change within the RP approach, againwith an application.

If time will also show how the approach can be extended to alife-cycle model with habit formation, and look at discrete choicemodels. All this is in the lecture notes anyway!

Summary so far....

I to test rationality through inequality restrictions, and

I to provide nonparametric estimates of bounds on counterfactualdemand responses.

Summary so far....

RP for Heterogeneous Consumers

I Assume every consumer is characterised by unobservedheterogeneity (ε) and responds to a given budget (p, x), with aunique, positive J−vector of demands

q = d(p, x , ε)

I As we noted one key drawback has been the (additive) separabilityof ε assumed in empirical specifications.

I in the non-separable case we will assume conditions onpreferences that ensure invertibility in ε,

I with J > 2, we will look at some new results on multiple goodswith nonseparable heterogeneity.

I for J = 2, invertibility is equivalent to monotonicity in unobservedheterogeneity ε

for example: I

q = d(p, x , ε)

for example: I

q = d(p, x , ε)

for example: I

q = d(p, x , ε)

for example: I

q = d(p, x , ε)

for example: I

q = d(p, x , ε)

for example: I

Invertible Preferences, J=2

For example, if preferences take the form:

U ti (q1i , q0i ) = v (q1i , q0i ) + w(q1i , εi )

preference heterogeneity εiw strictly increasing and concave with positive cross derivativeguarantees q1 is invertible in ε.

Note that RP consistent responses to price and income changes willbe represented by a shift in the distribution of demands.

We will assume baseline demands are monotonic in scalar unobservedheterogeneity so that quantile demands, conditional on x income andprice regime, identify individual demands.

Figure 2a: The distribution of heterogeneous consumers

Distribution of consumer tastes in a market:

q(x,ε)

Figure 2b: Monotonicity and rank preserving changes

q(x ,ε)

Figure 2c: The quantile expansion path

q(x,ε)

- > Quantile structural function - quantile Engel curve.

Nonseparable Demand

Consider the identification and estimation of demands

q (t) = d(x(t), t, ε), t = 1, ...,T ,

where the demand function in any price regime p (t) is exactly thestochastic expansion path (Engel curve) for prices for market t.

For the case with scalar heterogeneity, J = 2, the conditions for in εinvertibility correspond to monotonicity. For this case:

ε ∈ R,d(x(t), t, ε) = (d1(x(t), t, ε), d2(x(t), t, ε))

we make the following assumptions:

A 3.1: The variable x (t) has bounded support, x (t) ∈ X = [a, b] for−∞ < a < b < +∞, and is independent of ε ∼ U [0, 1], for now.A 3.2: The demand function d1 (x , t, ε) is invertible in ε and is continuouslydifferentiable in (x , ε).

Nonseparable Demand

Identification Result: d1(x , t, τ) is identified as the τth quantile of q1|x(t):

d1 (x , t, τ) = F−1q1(t)|x (t) (τ|x) .

Thus, we can employ standard quantile regression techniques to estimate d1.

Letρτ (y) = (I {y < 0} − τ) y , τ ∈ [0, 1] ,

be the check function used in quantile estimation. The budget constraintdefines the path for d2. We let D be the set of feasible demand functions,

D ={d ≥ 0 : d1 ∈ D1, d2 (x , t, τ) =

x − p1 (t) d1 (x , t, ε (t))p2 (t)

Estimation

Let (qi (t) , xi (t)), i = 1, ..., n, t = 1, ...,T , be i.i.d. observations from ademand system, qi (t) ∈ R2.Then estimate d (t, ·, τ) by

d (·, t, τ) = arg mindn∈Dn

∑i=1

ρτ (q1i (t)− d1n (xi (t))) , t = 1, ...,T ,

where Dn is a sieve space (Dn → D as n→ ∞).

Let Bi (t) = (Bk (xi (t)) : k ∈ Kn) ∈ R |Kn | denote basis functionsspanning the sieve Dn .

Then d1 (x , t, τ) = ∑k∈Kn πk (t, τ)Bk (x), where πk (t, τ) is astandard linear quantile regression estimator:

π (t, τ) = arg minπ∈R|Kn |

∑i=1

(q1i (t)− π′Bi (t)

), t = 1, ...,T .

BKM (2014) derive rates and asymptotic distribution of the sieve estimator.

RP-Restricted Estimation

There is no particular reason why estimated expansion paths for a sequenceof prices (markets) t = 1, ...,T satisfies RP for any type ε. To impose theRP restrictions, we simply define the constrained function set as:DTC = D

T ∩ {d (·, ·, τ) satisfies RP} .Define the constrained estimator by:

{dC (·, t, τ)}Tt=1

= arg min{dn(·,t ,τ)}Tt=1∈DTC

∑t=1

∑i=1

ρτ (q1,i (t)− d1,n (t, xi (t))) , τ ∈ [0, 1] .

Since RP imposes restrictions across t, the above estimation problem can nolonger be split up into T individual sub problems as in the unconstrainedcase. Adapting results on nonparametric estimation under shape constraints,show the constrained sieve estimator dC converges with the same rate as d .BKM (2014) demonstrate that as n→ ∞, the unrestricted estimator, d ,satisfies RP almost surely. Conclude that dC is asymptotically equivalent d ,and all the asymptotic properties of d are inherited by dC .Richard Blundell () Consumer Behaviour & Revealed Preference Short Course November 2017 46 / 89

Estimating Counterfactual Demand Bounds

For new budget (p0, x0) define the estimated income levelsx = (x (1) , ..., x (T )) as the solutions to

p′0dC (x (t) , t, τ) = x0, t = 1, ...,T ,

the support set estimator is Sp0,x0 = {q ∈ Bp0,x0 |x−Pq ≤ 0}.A valid confidence set can be constructed for the demand bounds, and isakin to the result found in, for example, CHT’s Theorem 5.2. Use modifiedbootstrap from Bugni (2009, 2010), Andrews and Soares (2010).We would also like to test for whether the consumers in the sample arerational (i.e. obey the RP restrictions).

Idea: Compute unrestricted demand estimates, and see how far they arefrom satisfying RP restrictions.Measure discrepancies between a given alternative set of demands,q = (q (1) , ...,q (T )) ∈ R2T , and q0 by:

MDn ({q (t)} |p0, x0)

∑t=1(q (t)− q0 (t,p0, x0))′Wt (q (t)− q0 (t,p0, x0)) ,

where Wt ∈ RJ×J is some weighting matrix.Still in development....

Estimating Counterfactual Demand Bounds

For new budget (p0, x0) define the estimated income levelsx = (x (1) , ..., x (T )) as the solutions to

p′0dC (x (t) , t, τ) = x0, t = 1, ...,T ,

the support set estimator is Sp0,x0 = {q ∈ Bp0,x0 |x−Pq ≤ 0}.A valid confidence set can be constructed for the demand bounds, and isakin to the result found in, for example, CHT’s Theorem 5.2. Use modifiedbootstrap from Bugni (2009, 2010), Andrews and Soares (2010).We would also like to test for whether the consumers in the sample arerational (i.e. obey the RP restrictions).Idea: Compute unrestricted demand estimates, and see how far they arefrom satisfying RP restrictions.Measure discrepancies between a given alternative set of demands,q = (q (1) , ...,q (T )) ∈ R2T , and q0 by:

MDn ({q (t)} |p0, x0)

∑t=1(q (t)− q0 (t,p0, x0))′Wt (q (t)− q0 (t,p0, x0)) ,

where Wt ∈ RJ×J is some weighting matrix.Still in development....

A sub-population from the UK FES diary records

Couples with two children from SE England

7 relative price changes

Couples with one child 1,421 and 1,906 observations per year.

Analyse spending on food and other non-durables.

Engel curve distribution for food in one market

3 3.5 4 4.5 5 5.5

0.20.2

0.40.4

log tota l ex p.

dens itymedian

Total Expenditure (Budget) Distribution

3 3.5 4 4.5 5 5.50

logtotal exp.

itykernel est.Normal est.

Estimation

I In the estimation, use a penalised quantile sieve estimator for theexpansion paths.

Show that the support set estimator inherits the (sup-norm)convergence rate of the underlying quantile sieve estimator.

Also how a valid confidence set can be constructed for the demandbounds, adapting moment inequality arguments in Chernozhukov,Hong and Tamer (2007).

Use these results to develop a test of the RP inequalities.

Use 3rd order pol. spline with 5 knots

RP restrictions imposed at 100 x-points over the empirical support xacross markets.

Estimation

Figure 4a. Unrestricted Quantile Expansion Paths: Food

3.8 4 4.2 4.4 4.6 4.8 5 5.2

logtotal exp.

.τ = 0.1τ = 0.5τ = 0.995% CIs

Figure 4b. RP Restricted Quantile Expansion Paths: Food

3.8 4 4.2 4.4 4.6 4.8 5 5.2

logtotal exp.

.τ = 0.1τ = 0.5τ = 0.995% CIs

Figure 5a: Quantile Counterfactual Demand Bounds atMedian Income and Median Heterogeneity

0.92 0.94 0.96 0.98 1 1.020

price, food

estimate95% confidence interval

Figure 5b: Estimated Counterfactual Demand Bounds asMore Markets are Added

Notes on the Estimates

Note the ‘local’nature of the analysis - the bounds vary with income,heterogeneity and the number of markets

Demand (e-)bounds (support sets) are defined at the quantiles of x and ε

tightest bounds given information and RP.show how vary with income and heterogeneity

To account for the endogeneity of x we can utilize IV quantile estimatorsdeveloped in Chen and Pouzo (2009) and Chernozhukov, Imbens and Newey(2007). The basic results remain valid in the quantile demand case exceptthat the convergence rate stated there has to be replaced by that obtainedin Chen and Pouzo (2009) or Chernozhukov, Imbens and Newey (2007).

Alternatively, use the control function approach taken in Imbens and Newey(2009) to recover the QSF. They use this data and the exact sameinstrument. Specify

ln x = π(z, v)

where π is monotonic in v , z are a set of instrumental variables.

The Slutsky Inequality

When prices and demand are continuous, the RP conditions for a singlegood become equivalent to the Slutsky inequality shape restriction on thesingle good demand (normalised to the outside good)

q1= d1(p, x , ε).

Blundell, Horowitz and Parey (2016) make the monotonicity assumption onε, and impose the Slutsky condition on the nonparametric estimate of theconditional quantile function.

The constrained estimator is obtained by solving a nonparametric quantileestimation problem subject to the Slutsky condition for all (p, x).

This problem has unaccountably many constraints. They replace thecontinuum of constraints by a discrete set, imposing the restriction on a grid.

Also implement an exogeneity test for prices and develop an IV estimator.

The National Household Travel Survey

Apply to gasoline demand in the US National Household Travel Survey(2001) - a household-level survey that was conducted by telephone andcomplemented by travel diaries and odometer readings.

To minimize heterogeneity, restrict the sample to a specific set ofdemographics.

Take vehicle ownership as given and do not investigate how changes ingasoline prices affect vehicle purchases or ownership.

The resulting sample contains 5,254 observations.

As an instrument for gasoline price we use distance from Gulf supply point.

The Shape Restricted Demand Curve (Median Demand atMedian Income)

The Shape Restricted Demand Curve (Median Demand atLow Income)

Findings

Imposing the Slutsky restriction on an otherwise fully nonparametricestimate of the demand function produces well-behaved estimates of thedemand function, avoiding arbitrary and possibly incorrect parametric orsemiparametric restrictions.The Slutsky constrained nonparametric estimates revealed features of thedemand function that are not present in simple parametric models, especiallyon price responses across the income distribution.

In recent work we note we do not observe the true transactions price.Instead, we observe a local (county) average price p that is related to p∗ by

p = p∗ + ζ

where ξ is an unobserved random variable.The resulting errors in variables are called “Berkson errors”and are commonin economics data - the opposite of classical errors in variables.Show this can produce important biases in the quantile estimation ofdemands but these are limited once the Slutsky inequality condition isimposed.

Findings

Imposing the Slutsky restriction on an otherwise fully nonparametricestimate of the demand function produces well-behaved estimates of thedemand function, avoiding arbitrary and possibly incorrect parametric orsemiparametric restrictions.The Slutsky constrained nonparametric estimates revealed features of thedemand function that are not present in simple parametric models, especiallyon price responses across the income distribution.In recent work we note we do not observe the true transactions price.Instead, we observe a local (county) average price p that is related to p∗ by

p = p∗ + ζ

where ξ is an unobserved random variable.The resulting errors in variables are called “Berkson errors”and are commonin economics data - the opposite of classical errors in variables.Show this can produce important biases in the quantile estimation ofdemands but these are limited once the Slutsky inequality condition isimposed.Richard Blundell () Consumer Behaviour & Revealed Preference Short Course November 2017 61 / 89

Multiple Goods and Many Errors

Multiple goods bring the full power of transitivity (symmetry) but, togetherwith non-separable heterogeneity also raise additional invertibility,identification and estimation issues, and nonseparable heterogeneity is‘essential’in multiple good demand models.

Generalise the previous many good example of BBC (2008) to avoid the useof average local demands. This requires invertibility of demand, see Matzkin(2007, 2010), Beckert and Blundell (2008), Berry, Gandhi, and Haile (2013).

The idea is to introduce variables Z that are correlated with unobservedheterogeneity ε. BKM (2017) limit the dimensionality of the unobservedheterogeneity, and focus on individual demands. Averages over asubpopulation are investigated in Hausman and Newey (2013) andBlomquist and Newey (2013). Related to the random coeffi cients models ofLewbel and Pendakur (2013).

Dette, H., S. Hoderlein, and N. Nuemeyer (2011) consider integrabilityconditions without invertibility but not necessarily for a particular individual.

Consider a demand system where G = J − 1 unobservable variables canenter in nonlinear, nonadditive ways.

q1 = d1 (p, x , ε1, ..., εG )q2 = d2 (p, x , ε1, ..., εG )

· · ·qG = dG (p, x , ε1, ..., εG )

where the vector of unobserved heterogeneity (tastes) (ε1, ..., εG ) isindependent of (p, x) conditional on Z .We make an invertibility assumption

ε1 = r1 (q1, ..., qG ,p, x)ε2 = r2 (q1, ..., qG ,p, x)

· · ·εG = rG (q1, ..., qG ,p, x)

In the application, ε = r (q,p, x) satisfies the Revealed Preferencerestrictions.Richard Blundell () Consumer Behaviour & Revealed Preference Short Course November 2017 63 / 89

Berry and Benkard (2006) and Matzkin (2007, 2008) note that withoutfurther restrictions the system is not identified. Our solution is based onvariables that are excluded from the functions of interest. Showidentification when:

A unimodal restriction with respect to Z on the conditional density of thevector of unobserved heterogeneity.

Develop methods to estimate the value of the vector of unobserved tastes ofeach consumer and the demand function of each consumer.

All methods are constructive and the estimators are shown to be consistentand asymptotically normal.

Assumption M : For some invertible H and given ε, there exists unique zsuch that

∂fε|Z=z (ε)

∂z= 0 <=> ε = H (z)

which requires: dim(z) =dim(ε) and fε,Z differentiable at z

An example:

ε = H (g(z) + η) ; η independent of Z ; mode of η known

A special case of which is:

ε = η − z ; η independent of Z ; mode of η known

as used in Matzkin (2007).

BKM are able to show identification of finite changes in budgets for anyindividual defined by a particular ε.

BBC (2008) application from the Family Expenditure Survey in the UK

food share and services share as functions of log(expenditure) and twounobserved tastes

z1 = family size calculated using equivalence scalesz2 = cohort, adjusted by education, of head of household

To disentangle the effects of price and preference change I want tolook at formalising the idea of taste change within the RP approach

If no rejection, set identification of objects of interestRationalisation with ‘well behaved’stable preferencesIf rejection, allow for taste change

Investigate the degree of ‘taste change’for tobacco and other ‘bads’

Address a specific question: How much of the fall in tobaccoconsumption in the UK was due to a rise in the relative price oftobacco and how much can be attributed to taste change?

Aim to inform policy on the balance between information/healthcampaigns and tax reform.

ABBC (2017) also consider how tastes evolve across differenteducation strata. Do tastes change differentially across educationgroups?

If no rejection, set identification of objects of interest

Rationalisation with ‘well behaved’stable preferencesIf rejection, allow for taste change

If no rejection, set identification of objects of interestRationalisation with ‘well behaved’stable preferences

If rejection, allow for taste change

Taste changes and pricesUK Budget shares for Tobacco: Quantiles

1 9 8 0 1 9 8 5 1 9 9 0 1 9 9 5 2 0 0 00

0 .0 2

0 .0 4

0 .0 6

0 .0 8

0 .1 2

0 .1 4

0 .1 6

T im e

0 .5 5 Q u a n ti le

1 9 8 0 1 9 8 5 1 9 9 0 1 9 9 5 2 0 0 00

0 .0 5

0 .1 5

T im e

(a) Low Education (b) High Education

Taste Change

Consumer i’s maximisation problem can be expressed as:

maxqui (q, αit ) subject to p

′q = x

where q ∈ RK+ denotes the demanded quantity bundle, p ∈ RK

denotes the (exogenous) price vector faced by consumer i and x givestotal expenditure.

αit is a potentially infinite-dimensional parameter that indexesconsumer i’s tastes at time t. This allows for taste change for anygiven consumer across time.

We also allow for unobserved permanent heterogeneity acrossconsumers.

Using this framework we derive RP inequality conditions thatincorporate minimal perturbations to individual preferences to accountfor taste change.

Taste Change

′q = x

Taste Change

′q = x

Marginal utility (MU) perturbations

MU perturbations represent a simple way to incorporate tastevariation: McFadden & Fosgerau, 2012; Brown & Matzkin, 1998,represent taste heterogeneity as a linear perturbation to a base utilityfunction.

Characterising taste change in this way yields the temporal series ofutility functions:

ui (q, αit ) = vi (q) + αi ′t q, where αit ∈ RK .

Under this specification, αi ,kt can be interpreted as the taste shift inthe marginal utility of good k at time t for individual i .

The theorems below imply this specification is not at all restrictive.

Marginal utility (MU) perturbations

MU perturbations represent a simple way to incorporate tastevariation: McFadden & Fosgerau, 2012; Brown & Matzkin, 1998,represent taste heterogeneity as a linear perturbation to a base utilityfunction.

Characterising taste change in this way yields the temporal series ofutility functions:

ui (q, αit ) = vi (q) + αi ′t q, where αit ∈ RK .

Under this specification, αi ,kt can be interpreted as the taste shift inthe marginal utility of good k at time t for individual i .

The theorems below imply this specification is not at all restrictive.

Afriat conditions

For individual i we seek the Afriat inequalities that would allow us torationalise observed prices

{p1, ...pT

}and quantities

{q1, ...qT

We can ‘good 1 taste rationalise’the observed prices and quantities ifthere is a function v (q) and scalars {α1, α2, ...αT } such that:

v(qt)+ αtqt1 ≥ ψ (q) + αtq1

for all q such that ptq ≤ ptqt .

Afriat conditions

Theorem: The following statements are equivalent:

1. Individual observed choice behaviour, {pt ,qt}t=1,...,T , can be good-1rationalised by the set of taste shifters {αt}t=1,...,T .2. One can find sets {vt}t=1,...,T , {αt}t=1,...,T and {λt}t=1,...,T withλt > 0 for all t = 1, ...,T , such that there exists a non-empty solution setto the following inequalities:

(v (qt )− v (qs )) + αt (qt1 − qs1) ≤ λt (pt )′ (qt − qs )

αt ≤ λtpt

These inequalities are a simple extension of Afriat (1967).

When they hold there exists a well-behaved base utility function and aseries of taste shifters on good-1 that perfectly rationalise observedbehaviour.

Afriat conditions

Theorem: The following statements are equivalent:

1. Individual observed choice behaviour, {pt ,qt}t=1,...,T , can be good-1rationalised by the set of taste shifters {αt}t=1,...,T .2. One can find sets {vt}t=1,...,T , {αt}t=1,...,T and {λt}t=1,...,T withλt > 0 for all t = 1, ...,T , such that there exists a non-empty solution setto the following inequalities:

(v (qt )− v (qs )) + αt (qt1 − qs1) ≤ λt (pt )′ (qt − qs )

αt ≤ λtpt

These inequalities are a simple extension of Afriat (1967).

When they hold there exists a well-behaved base utility function and aseries of taste shifters on good-1 that perfectly rationalise observedbehaviour.

A surprising result

We can then show, under mild assumptions on the characteristics ofavailable choice data, that we can always find a pattern of tasteshifters on a single good that are suffi cient to rationalise any finitetime series of prices and quantities:

Definition: There is ‘perfect intertemporal variation’(PIV) in good 1if qt1 6= qs1 for all t 6= s = 1, ...,T .Theorem: Given observed choice behaviour, {pt ,qt} for t = 1, ...Twhere good-1 exhibits PIV, one can always find a set {vt , αt ,λt} withλt > 0 for all t = 1, ...,T , that satisfy the Afriat inequalities.

PIV is suffi cient for rationalisation but not necessary.

A surprising result

We can then show, under mild assumptions on the characteristics ofavailable choice data, that we can always find a pattern of tasteshifters on a single good that are suffi cient to rationalise any finitetime series of prices and quantities:

Definition: There is ‘perfect intertemporal variation’(PIV) in good 1if qt1 6= qs1 for all t 6= s = 1, ...,T .Theorem: Given observed choice behaviour, {pt ,qt} for t = 1, ...Twhere good-1 exhibits PIV, one can always find a set {vt , αt ,λt} withλt > 0 for all t = 1, ...,T , that satisfy the Afriat inequalities.

PIV is suffi cient for rationalisation but not necessary.

Taste changes as price adjustments

We can reinterpret the rationalisability question as a ‘missing priceproblem’.

We can find scalars {v1, ...vT }, positive scalars {λ1, ...λT }, and aweakly positive taste-adjusted price vector, {pt}t=1,..,T , such that

v(qt)− v (qs ) ≥ λt

(pt)′ (qt − qs)

wherept =

[pt1 − αt/λt ,pt¬1

We refer to αt/λt as the taste wedge.

The change in demand due to a positive taste change for good 1(αt > 0) can be viewed as a price reduction in the price of good 1.

This provides a link between two of the levers (taxes and information)available to governments.

Recovering taste change perturbations

Given the no rejection result, we can always find a non-empty set ofscalars that satisfy the Afriat conditions.Pick out values {vt , αt ,λt}t=1,...T that solve:

∑t=2

α2t subject to the Afriat inequalities

This a quadratic-linear program.Minimizing the sum of squared α’s subject to the set of RPinequalities ensures that the recovered pattern of taste perturbationsare suffi cient to rationalise observed choice behaviour.With α1 = 0, we interpret {αt}t=2,...,T as the minimal rationalisingmarginal utility perturbations to good-1 relative to preferences att = 1.Can also impose more structure on the evolution of taste change overtime. For example, monotonicity: αt+1 ≤ αt .

Empirical Strategy

Our empirical analysis uses data drawn from the U.K. FamilyExpenditure Survey (FES) between 1980 and 2000.

The FES records detailed expenditure and demographic informationfor 7,000 households each year.

It is not panel data so we follow birth-cohorts of individuals stratifiedby education level.

Estimation

To operationalise we estimate censored quantile expansion paths ateach price regime (see Chernozhukov, Fernandez-Val and Kowalski(2010)) subject the RP inequalities.

Separately by birth cohort and by education group E i ∈ {L,H} .

We use a quantile control function approach to correct for theendogeneity of total expenditure.

We recover shifts in the distribution of demands and ask what are theminimal perturbations to tastes that maintain the RP inequalities ateach particular quantile.

Estimation

Results

Minimal virtual prices along each birth cohort’s SMP path τthquantile and education group E are recovered as:

πE ,τt = p1t −αE ,τt

λE ,τt

The "taste wedge", αE ,τt /λE ,τt represents the change in the marginal

willingness to pay for tobacco relative to base tastes.

We find:1 Some degree of taste variation is necessary to rationalise observedbehaviour.

2 There are significant differences in the path of systematic tastechange between education cohorts for light and moderate smokers.

3 The taste change trajectories for light and moderate smokers in thehigh education cohort are similar.

4 Education is irrelevant for explaining the evolution of virtual pricesamongst heavy smokers.

Results

λE ,τt

Results

λE ,τt

Results

λE ,τt

Results

λE ,τt

4 Education is irrelevant for explaining the evolution of virtual pricesamongst heavy smokers.Richard Blundell () Consumer Behaviour & Revealed Preference Short Course November 2017 77 / 89

Taste wedges for light smokers

Taste wedges for medium smokers

Taste wedges for heavy smokers

Results: relaxing separability through conditional demands

Weak separability with alcohol consumption is a strong assumption.Alcohol is often thought to be complementary with tobaccoconsumption.

To relax this weak separability assumption we re-run our quadraticprogramming procedure on quantile demands that are estimatedconditional on alcohol consumption.

We partition the set of observations into "light" and "heavy" drinkersdepending on whether an individual is below or above the medianbudget share for alcohol.

The significant difference by education group in the evolution tastechange for light and moderate smokers is robust to non-separability.

95% confidence intervals on virtual prices and the taste wedge aredisjoint across education groups for all cohorts except for the "heavysmoking"-"heavy drinking" group. Effective tastes for this groupevolved very little for both education groups.

Taste Wedge Results: Conditional Quantiles (ModerateSmoker)

1 9 8 0 1 9 8 5 1 9 9 0 1 9 9 5 2 0 0 0 1 2

L o wH ig h

1 9 8 0 1 9 8 5 1 9 9 0 1 9 9 5 2 0 0 0 1 2

T im e

H ig h

Light Drinker Heavy Drinker

Taste Wedge Results: Conditional Quantiles (HeavySmoker)

1 9 8 0 1 9 8 5 1 9 9 0 1 9 9 5 2 0 0 0 1 2

T im e

L o wH ig h

1 9 8 0 1 9 8 5 1 9 9 0 1 9 9 5 2 0 0 0 1 2

L o wH ig h

Light Drinker Heavy Drinker

Characterising Taste Change

In this final part of the lecture we have shown how to develop anempirical framework for characterising taste change that recovers theminimal intertemporal (and interpersonal) taste heterogeneityrequired to rationalise observed choices.

A censored quantile approach was used to allow for unobservedheterogeneity and censoring.

Non-separability between tobacco and alcohol consumption wasincorporated using a conditional (quantile) demand analysis.

Future work will use intertemporal RP conditions to recover the pathof λt .

Systematic taste change was required to rationalise the distribution ofdemands in our expenditure survey data. Statistically significanteducational differences in the marginal willingness to pay for tobaccowere recovered; more highly educated cohorts experienced a greatershift in their effective tastes away from tobacco.

Overall Summary

I to test rationality, andI to improve the performance of nonparametric estimates of demandresponses with unobserved heterogeneity

I New (empirical) insights provided about the distribution of priceresponsiveness by unobserved heterogeneity, income and otherobserved characteristics of consumers.

Formalise the notion of taste change within the RP approach.

I For example, evidence that tobacco consumption by low educationhouseholds can be largely rationalised by relative prices whereas tastechanges are key in the decline for higher educated households.

Extend to a life-cycle model with habit formation.

Overall Summary

I to test rationality, and

I to improve the performance of nonparametric estimates of demandresponses with unobserved heterogeneity

Overall Summary

Extra Slide 1: Life-cycle Planning and Habits

Allow for short memory in tobacco consumption such that the baseutility function depends on lagged quantity of good 1:

υt = ψ(q, q−11

)+ µtq1

Following Browning (1989) and Crawford (2010), embed this felicityfunction in a standard lifecycle planning framework.

max{qt}t=1,...,T

∑t=1

βt−1{

ψ(qt , qt−11

)+ µtq

∑t=1

ρ′tqt = A0

for discounted prices ρt .

Extra Slide 1: Life-cycle Planning and Habits

Allow for short memory in tobacco consumption such that the baseutility function depends on lagged quantity of good 1:

υt = ψ(q, q−11

)+ µtq1

Following Browning (1989) and Crawford (2010), embed this felicityfunction in a standard lifecycle planning framework.

max{qt}t=1,...,T

∑t=1

βt−1{

ψ(qt , qt−11

)+ µtq

∑t=1

ρ′tqt = A0

for discounted prices ρt .

Extra Slides 2: Stochastic Revealed Preference and RUM

The Random Utility Model framework:

Let u be a random utility function (unobserved heterogeneity) for consumersfacing prices pt and budget set Bt . Assume repeated cross section data.We observe (a sample analog of) choice probabilities and ask: Can theobservable choice probabilities be rationalized as an outcome of the RUM?As before this will then allow a prediction/ counterfactual analysis only usingtheoretical restrictions implied by the RUM?

Follow recent work by Kitamura and Stoye (2017) who develop thetheoretical background on the Axiom of Stochastic Revealed Preference(ASRP) from McFadden and Richter (1991) [MR], McFadden (2005).

Also reference Manski (2012) who develops a nonparametric analysis oflabor supply with revealed preference.

SRP and RUM

McFadden 2005 shows, without further assumptions on individual preferences(and unobserved heterogeneity) it is suffi cient to consider finite partitions of theseparate non-intersecting sections of the budget constraint. This is also used inHoderlein and Stoye.

Note that with discrete choice

Cannot use point demands and Engel curves.How choices are made within a partition does not provide additionalinformation.The choice space is the collection of partitions.

SRP and RUM

To implement the MR theory of Stochastic Rationality, Kitamura and Stoye(2017)[KS] define u∗ to be a realization of u and define the vector a(u∗) as achoice pattern over the choice set implied by u∗.

Note that certain patterns are not allowed due to SARP, so create a matrixA of all (H) such valid vectors.The matrix A plays a key role, embodying the restrictions due to rationality.KS discuss algorithms to obtain the matrix A.

If there exists a probability vector v for the H types such that

Av = π

then the choice probabilities are stochastically rational i.e. satisfy the Axiomof Revealed Stochastic Preference [ARSP].

Various equivalent statements for ARSP have been noted by MR, KS chooseto work directly with Av = π. Use this to develop a nonparametericapproach to test the stochastic rationality hypothesis.

Consumer Behaviour and Revealed Preference - …uctp39a/blundell-short-course-2017-lecture3.pdf ·...

Documents

Transcript of Consumer Behaviour and Revealed Preference - …uctp39a/blundell-short-course-2017-lecture3.pdf ·...